For years, the conversation around AI in Singapore’s financial sector felt a bit like a high-level philosophy seminar. We talked about “ethics,” we debated “fairness,” and we looked at the FEAT principles as a North Star. But as of late February 2026, the philosophy seminar is officially over. The operational manual has arrived.
The Monetary Authority of Singapore (MAS) has concluded its consultation on the new Guidelines on AI Risk Management (AIRG), and for every Financial Institution (FI) in the Lion City, the 12-month countdown has begun. This isn’t just another layer of compliance; it is a fundamental shift in how we build, deploy, and supervise the “digital brains” running our financial systems.
As someone who has advised VPs of Risk and Directors of Innovation across the region, I can tell you that the mood in the boardroom has shifted from curiosity to a very focused kind of urgency. The question is no longer “What can AI do for us?” but “How do we prove to MAS that we can control it?”
From Ethics to Supervisory Reality
The AIRG represents a maturation of the MAS’s approach. If FEAT (Fairness, Ethics, Accountability, and Transparency) provided the what, the AIRG provides the how. MAS is no longer just asking you to be ethical; they are expecting you to be systematically robust.
The scope is intentionally broad. It covers everything from the traditional machine learning models used in credit scoring to the latest generative AI tools and the emerging class of agentic AI systems that can plan and act autonomously.
Frankly, the biggest shock for many of my clients isn’t the technical requirement—it’s the accountability mandate. Under the new guidelines, the Board and Senior Management are explicitly held responsible for AI risk. You cannot delegate “AI oversight” to a junior data scientist in the basement. It is now a top-tier corporate governance issue.
Pillar 1: The AI Inventory and Risk Materiality
The first step in the 12-month countdown is arguably the most difficult: you have to know what you have. MAS now expects FIs to maintain a comprehensive AI Inventory.
I recently spoke with a Director of Risk at a mid-sized regional bank who thought they had three AI use cases. After a thorough audit, we found fourteen. Many were “shadow AI” projects—department-level initiatives using third-party APIs that hadn’t been through a formal risk assessment.
Once you have your inventory, you must perform a Risk Materiality Assessment based on three dimensions:
- Impact: How much could this model lose, or how many customers could it unfairly disadvantage?
- Complexity: Do we actually understand how this “black box” makes decisions?
- Reliance: What happens if this AI goes offline or starts hallucinating?
The bottom line is that MAS is taking a proportional approach. They don’t expect the same level of rigour for a chatbot that summarises meeting notes as they do for a model that determines a small business’s creditworthiness.
Pillar 2: The “Agentic” Challenge
One of the most forward-looking aspects of the 2026 landscape is the integration of the Model AI Governance Framework for Agentic AI, launched in January.
In the financial world, we are moving rapidly from “Advisory AI” (the system gives a human a recommendation) to “Agentic AI” (the system executes the trade or approves the loan autonomously). MAS is acutely aware of the “black box” risk here.
The new guidelines mandate strict “risk bounding” for autonomous agents. This means you must have hard-coded limits on what an agent can do without human intervention. I remember advising a VP of Operations at a wealth management firm who wanted to let an AI agent “optimise” client portfolios in real-time. My advice was simple: “Optimisation is fine, but execution needs a ‘dead-man’s switch.’” Under AIRG, that switch isn’t just good practice; it’s a requirement.
Pillar 3: End-to-End Lifecycle Controls
The AIRG doesn’t just care about the output of the AI; it cares about the entire lifecycle—from data ingestion to retirement. This is a significant departure from the old way of doing things, where we often focused only on the “final” model. Now, every stage of the process must be documented and auditable.
This means your data management needs to be impeccable. Where did the training data come from? Is it biased? How do you ensure that “poisoned” data hasn’t been introduced to a generative AI model? You need to have clear protocols for data lineage and version control. If a model starts making skewed credit decisions six months from now, you must be able to trace that back to the specific dataset that caused the shift.
FIs are now being pushed toward using tools like AI Verify, Singapore’s open-source governance testing framework. By mapping AI Verify to international standards like ISO/IEC 42001, MAS is giving FIs a way to provide technical assurance that their models are doing what they claim to do. But technical assurance is only half the battle. You also need a robust “Human-in-the-Loop” strategy. MAS is very clear: autonomy does not mean an absence of accountability. If the AI makes a mistake, a human must be able to explain why it happened and, more importantly, how to fix it. This requires a level of “explainability” that many current models simply don’t possess, forcing a rethink of model selection toward more transparent architectures.
The Role of AI Red Teaming: Testing the Breaking Point
One of the most critical new requirements under the 2026 guidelines is the mandate for adversarial testing, or “AI Red Teaming.” In the past, red teaming was the domain of cybersecurity. Today, it is a core component of AI governance.
FIs must now actively try to break their own models. This involves simulating attacks that attempt to bypass safety filters, extract sensitive training data, or manipulate the model’s output through “prompt injection.” I remember a project with a regional insurance provider where our red team was able to convince their “secure” customer service agent to reveal confidential internal pricing structures just by using a specific sequence of ambiguous questions.
Under AIRG, these vulnerabilities aren’t just bugs; they are regulatory failures. FIs must show that they have conducted these tests and, more importantly, that they have implemented the necessary guardrails to prevent such exploits in a production environment. This isn’t a one-time test; it must be a continuous part of the AI lifecycle, especially for generative models that are constantly being updated with new data.
The 12-Month Roadmap: A Practical Guide for the C-Suite
If you are a VP or a Director in a Singaporean FI today, your roadmap for the next twelve months should look like this:
Month 1-3: Discovery and Audit
Complete your AI Inventory. Don’t just ask IT; ask Marketing, ask HR, ask your Customer Service leads. Map every model to its risk materiality. This is also the time to identify any third-party AI dependencies. If your mortgage processing depends on an external AI vendor, their risk is now your risk.
Month 4-6: Governance Structuring
Establish your AI Committee. This shouldn’t just be data scientists. You need Risk, Legal, Compliance, and Business heads in the room. This committee needs to report directly to Senior Management. Their job is to define the “risk appetite” for AI—deciding, for example, exactly how much autonomy a trading agent is allowed to have before a human must hit “approve.”
Month 7-9: Technical Validation and Red Teaming
Begin using assurance frameworks like AI Verify for your high-risk models. Conduct your first formal AI Red Teaming exercises. If you’re using third-party AI, start demanding “Transparency Reports” from your vendors. If they can’t provide them, you may need to reconsider the partnership. This is the stage where you move from “policy” to “proof.”
Month 10-12: Training and Cultural Shift
This is the most overlooked part, yet it is arguably the most important. You need to train your staff on the new “human-in-the-loop” requirements. They need to understand that their role isn’t just to “use” the AI, but to “supervise” it. This requires a new kind of “AI Literacy” that goes beyond technical skills to include critical thinking and ethical judgement.
I’ve seen too many organisations treat this as a technical training exercise. It isn’t. It’s a cultural shift. Your VPs and Directors need to be comfortable challenging the output of an AI. They need to know that “the AI said so” is never an acceptable answer in a MAS audit. We need to move from a culture of “automation at all costs” to one of “accountable autonomy.” This means rewarding employees who catch AI errors and fostering an environment where questioning the algorithm is encouraged, not discouraged.
The Bottom Line: Compliance as a Competitive Moat
I know many in the industry view the AIRG as a burden. But I see it differently.
In a world where trust in financial systems is paramount, and where AI hallucinations are becoming a boardroom risk, having a “MAS-compliant” AI framework is a massive competitive advantage. It tells your customers, your investors, and your partners that you aren’t just playing with new toys—you are building a resilient, governed, and trustworthy digital future.
The 12-month countdown has started. The FIs that treat this as a “tick-the-box” exercise will find themselves struggling when the thematic reviews begin in 2027. But the ones that use this year to fundamentally redesign their AI architecture will be the leaders of the next decade.
The Lion City has set the standard. Now, it’s time for our financial institutions to meet it.